Skip to main content
Advertisement

< Back to Article

Fig 1.

Overall flowchart of our proposed training and learning methodology for the sickle RBC-dCNN classification model describing the four main steps, including an independent shape factor analysis.

More »

Fig 1 Expand

Fig 2.

Hierarchical RBC patch extraction algorithm workflow.

See also Figs 3 and 4 for details of each step.

More »

Fig 2 Expand

Fig 3.

Determining ROIs and RBC patch extraction based on information from the entropy statistical estimation and morphology operations.

(A) raw microscopy image. (B) Blocks with high entropy. (C) ROI mask image. (D) Detection of ROIs. (E) Boundaries of ROIs. (F) Single & “touching/overlapped”.

More »

Fig 3 Expand

Fig 4.

Workflow of RBC patch generation from ROIs with touching RBCs.

(A) ROI patch. (B) binary ROI mask image. (C) Euclidean distance transfrom result. (D) Probability map based on random walk method with seeds (green dots) obtained from distance transform result. (E) separated RBC binary mask. (F) segmented RBC outlines (red line). (G) bounding boxes of single RBCs. (H) The touching RBCs are separated into four single RBC patches.

More »

Fig 4 Expand

Fig 5.

Workflow of the size-invariant (100px*100 px) RBC patch normalization.

Steps 1-5 are described in the text.

More »

Fig 5 Expand

Fig 6.

Size-invariant RBC patch size normalization for eight different types of diseased RBCs (horizontal) and four different groups (vertical).

Images in the first row should be compared against the first column of “Discocytes”.

More »

Fig 6 Expand

Fig 7.

Architecture of deep RBC-dCNN for SCD RBC classification.

More »

Fig 7 Expand

Table 1.

Description of our experimental dataset from eight patients’ imaging data.

More »

Table 1 Expand

Fig 8.

Distribution of the number of images in five fractions corresponding to seven different patients.

More »

Fig 8 Expand

Fig 9.

Convergence studies on loss and train error with respect to the number of iterations and different learning rate settings.

More »

Fig 9 Expand

Fig 10.

Normalized confusion matrix results with respect to different number of iterations.

(A) max_iter = 30. (B) max_iter = 60.

More »

Fig 10 Expand

Table 2.

Comparisons of loss and train errors based on different iterations.

More »

Table 2 Expand

Fig 11.

Evaluation procedure for 5-fold cross validation.

More »

Fig 11 Expand

Table 3.

5-fold cross validation for 5 types of SCD RBC classification.

More »

Table 3 Expand

Fig 12.

ROC-AUC result for coarse 5 types of RBC classification based on “Exp_II” dataset by 5-fold cross validation.

More »

Fig 12 Expand

Fig 13.

Performance analysis for SCD RBC classification based on “Exp_II” dataset (coarse labeling).

(A) confusion matrix showing the detailed number of correctly classified RBC images and misclassified RBC images, (B) 5 statistic metrics for the 5 types of RBC category prediction results, (C) Specific F-score, precision and recall analysis for different RBC categories.

More »

Fig 13 Expand

Table 4.

5-fold cross validation for 8 types of SCD RBC classification.

More »

Table 4 Expand

Fig 14.

ROC-AUC result for refined 8 types of RBC classification based on “Exp_II” dataset by 5-fold cross validation.

More »

Fig 14 Expand

Fig 15.

Classification performance analysis for SCD RBC classification based on “Exp_II” dataset (refined labeling).

(A)confusion matrix showing the detailed number of correctly classified RBC images and misclassified RBC images. (B) 5 statistic metrics for the 8 RBC category prediction result. (C) Specific F-score, precision and recall analysis for different RBC categories.

More »

Fig 15 Expand

Fig 16.

Prediction result example for 8 types of RBCs.

More »

Fig 16 Expand

Fig 17.

Statistical quantification for the number of various types of RBCs in density fraction 4 (severe SCD).

Notice the significant heterogeneity of cell types even at the highest density fraction.

More »

Fig 17 Expand

Table 5.

Description of our experimental dataset for the coarse-labeled five categories and a new deoxygenated category (row 6).

More »

Table 5 Expand

Fig 18.

First row: (t = 0) Irreversibly sickled cells (ISCs) under normoxia i.e., Oxy state [O2]: 20%; Initiation of deoxygenation. Second row: (t = 45s) Deoxygenated ISCs i.e., DeOxy state [O2]: 2%; (see S4 Appendix methods on Oxy/DeOxy states).

More »

Fig 18 Expand

Fig 19.

Snapshots of feature maps for layer 5 including the deoxygenated category.

(A) original images in batch 1. (B) feature maps in 5th layer.

More »

Fig 19 Expand

Fig 20.

Learning hierarchical feature maps for layers 5, 6, 8, 10 for Dic+Ovl (Oxy), Ech (Oxy), El+Sk (Oxy), El+Sk (DeOxy).

More »

Fig 20 Expand

Fig 21.

Confusion matrix for the combined oxygenated and deoxygenated categories described in Table 5.

More »

Fig 21 Expand

Fig 22.

CSF and ESF shape factors estimation for Elongated, Oval and Discocytes RBC types.

More »

Fig 22 Expand